应用神经网络的图像分类矢量量化编码

Classified VQ Coding of Image Using Neural Network

矢量量化（ＶｅｃｔｏｒＱｕａｎｔｉｚａｔｉｏｎ）作为一种有效的图像数据压缩技术，越来越受到人们的重视。设计矢量量化器的经典算法ＬＢＧ算法，由于运算复杂，从而限制了矢量量化的实用性。本文讨论了应用神经网络实现的基于边缘特征分类的矢量量化技术。它是根据人的视觉系统对图像的边缘的敏感性，应用模式识别技术，在对图像编码前，以边缘为特征对图像内容分类，然后再对每类进行矢量量化。除特征提取是采用离散余弦变换（ＤＣＴ）外，图像的分类和矢量量化都是由神经网络完成的。实验结果表明，和单纯用神经网络直接进行矢量量化相比，应用这种技术的图像编码压缩比和译码图像质量都有明显的提高。

Recently the vector quantization(VQ) has received considerable interests as a powerful image data compression technique.The most popular algorithm for VQ codebook design has been the LBG.While the LBG algorithm and its variants have been widely studied,the practical application of VQ has been limitted because of the prohibitive amounts of computation associated with both the vector encoding and the codebook design stages.In this paper,we describe the use of self-organizing neural networks in the image classified VQ(CVQ) based on edge feature classification.The principle of this technique is:According to human visual system s sensitivity to image edges,using pattern recognition technique to classify the image data into several classes based on edge features,then encoding every class of data with VQ.All computation,except for the edge feature extraction which is performed by DCT,are carried out by neural networks.The experiment results show,compared to direct image VQ coding without classification,both the compression ratio and image quality are well improved.